Constrained Nonnegative Matrix Factorization With Adaptive Weight For Hyperspectral Image Unmixing

Spectral unmixing plays an important role in the application of hyperspectral images.Owing to the low spatial resolution characteristic of hyperspectral sensors and the complex distribution of materials,one pixel in hyperspectral image may contain some kinds of materials that is to say the “mixed pixels”.The presence of “mixed pixels” has brought many difficulties to the quantitative of the hyperspectral images.Hyperspectral unmixing aims to decompose the measurement spectrum of “mixed pixels” into endmember and their responding abundance.As a popular unsupervised method,non-negative matrix factorization(NMF)has received increasing attention in recent years.The thesis focuses on the research of the constrained NMF-based unmixing methods for hyperspectral images.The imbalanced data in the hyperspectral images may be ignored,resulting in a low unmixing accuracy.By making use of the clustering analysis methods,sparseness expression,manifold learning and Min Max modeling methods,this paper focuses on the research of the unmixing model and method based on the cluster-wise weighted NMF for hyperspectral images with imbalanced data,as well as the Min Max NMF with weight adaptive for hyperspectral images.The main research works and innovations are listed as follows:1.There may be a large difference in the number of samples related to different endmembers,which is mainly caused by the imbalanced nature of the target distribution.In previous studies,most NMF-based unmixing methods have neglected the imbalanced samples due to the statistical characteristic of NMF,which causes the corresponding endmembers to be estimated with lower accuracy.Regarding to above problems,a clusterwise weighted NMF for hyperspectral images unmixing with imbalanced data(CW-NMF)is proposed.First,the K-means clustering method is used to cluster the given hyperspectral dataset to obtain the clustering information of different endmembers.Second,the weight of the cluster included the different endmembers is calculated with this clustering information.With this method,the pixels that include imbalanced endmembers are assigned larger weight values,while the pixels mixed with majority endmembers are given smaller weight values.Finally,the weights are extended to the entire hyperspectral image dataset and weighted into the standard NMF model.Moreover,the proposed unmixing model has a good extensibility by incorporating the sparsity constraint of abundance and graph-based regularization,respectively.Synthetic and real data experiments demonstrate that the proposed method achieves better unmixing accuracy than the existing methods.2.The CW-NMF method uses the K-means clustering method to preprocess the data,and its unmixing results are affected by the initial clustering center.Aiming to solve the problem that CW-NMF method is confronted with the interferences from both poor robustness and noise sensitivity,the Min Max non-negative matrix factorization with adaptive weight for hyperspectral image unmixing is proposed.First,the Min Max K-means clustering method is used instead of the K-means clustering method to optimize the variance in each cluster.Secondly,clustering analysis are conducted on the hyperspectral images,and the clustering procedure can explicitly characterize the structure information of the given data,which partitions the data points into different classes.Similar spectral signatures often imply similar representations of the decomposition with respect to the given endmember matrix.The clustering-based approach reveals a great potential for exploiting and preserving the structure of the data.Therefore,the obtained clustering information is used to explicitly represent the structural information of the given data.Finally,the cluster optimization results and the structure information of the data are combined to represent the NMF model.Moreover,the weight matrix in this method are adaptive updated for each iteration during the unmix process.The results of synthetic and real data experiments have proved that the proposed algorithm not only improve the unmixing accuracy,but also has good robustness and stability.